Theory of wisdom

Compile | Bot

Source | Google Research Blog

Editor’s note: On Wednesday, Google introduced an experimental semantic parser for a natural language framework by scientists Michael Ringgaard and Rahul Gupta. SLING’s new tool offers some exciting ideas based on frame semantics, but Google is only half-talking. There are no papers, no case studies, and no evidence of model accuracy, so we can’t know what breakthroughs it will make in framing.

Here’s what Google said in its blog post:

Nowadays, most of the practical natural semantic understanding (NLU) still using analytical method of step by step, from the part of speech tagging to generate word vector to generate interdependence syntactic analysis to text analysis, although the modular approach helps training model, but it also brings a problem: the early stage of the cascading effects with mistakes, late directly affects the output and the final accuracy of the model.

SLING, a semantic parser for natural language frameworks recommended by Google today, addresses this problem by building a semantic framework diagram based on natural language text. After a statement is entered, the framework captures what the user is interested in and annotates it. The hierarchical steps are parallel, which Google says avoids “pipelining” pitfalls and unnecessary computations. SLING is based on a dedicated recurrent neural network (RNN) that incrementally compiles input text directly on a block diagram. Block diagrams are flexible enough to efficiently extract semantic tasks of interest to developers; SLING’s parser is word-only, so you can bypass some intermediates, such as dependency generation parsing.

SLING produced code that was fast to refine on RNN, with parsing speeds of over 2500 tokens/SEC on desktop-level cpus. This mainly relies on the following two designs:

  • An efficient and extensible framework storage implementation;

  • A JIT compiler.

It is implemented in C++ and can be downloaded from github for interested readers.

Framework semantic analysis

SLING’s linguistic theory is frame semantics, a branch of cognitive linguistics that uses semantic frameworks, semantic networks, rather than syntactic analysis, to understand the meaning of words. It regards a sentence as a semantic framework in which the names of relevant semantic roles are different depending on the semantic categories of the lexical elements.

In the following example, example sentences can be simply and roughly divided into three components for SLING to categorate: entities (people, places, events, etc.), metrics (dates, distances, etc.), and other concepts (verb like components, etc.) —

Many people now claim to have predicted Black Monday

People is a person, Black Monday is an event, and Predicted is a behavior that connects people and events (other concepts). Predict-01 can be used as a framework element alone. When interacting with a character, predict-01 uses ARG0 slot to indicate “who made the prediction” and ARG1 slot to indicate “what was predicted” when interacting with an event. The whole judgment process is based on semantic conceptual structure and is not refined for every verb.

Although this example is simple, this approach works particularly well in complex sentences involving reference, metaphor, and metonymy, based on existing studies of frame semantics. It is worth mentioning that it is suitable for developers who use corpora as models.

SLING

SLING trains RNN by optimizing the semantic framework.

As mentioned earlier, in SLING, the current level of analysis is parallel, so instead of manual word segmentation and pipelined feature extraction, it puts those steps into a hidden layer. Within the neural network, SLING uses an encoder-decoder architecture that primes the input word into a vector with a suffix (ES) attached to it based on the semantic features of the dictionary. The decoder also handles punctuation marks.

SLING then combines them with their respective historical features, analyzing the ARG slots that make up the meaning one by one to get the expected semantic framework of the sentence. This model has been trained in TensorFlow and DRAGNN.

SLING, as the video shows, analyzes word-by-word elements of the semantic framework and generates sentence structures based on their relationships. Once the last word is entered, the whole model is adjusted once again, feeding the developer the most complete semantic framework. It’s worth noting that some words are entered and SLING stores them in a buffer if their role in the frame isn’t immediately obvious, and then triggers a judgment from subsequent words. Without doing syntactic dependency generation, this approach is effective in catching relationship exceptions between framework elements.

Theory of intelligence review

Since the blog only introduces this idea and does not show specific training examples and model accuracy, it is temporarily impossible to judge SLING’s reliability. And Google’s secretive, unsubstantiated approach also really makes people angry. SLING’s chances of having a full-fledged framework are slim to none based on current data on frame semantics, though if it did, Google Translate would be a lot better than it is now in the first place.

For those who are interested in SLING, or in the semantics of frames, look at the core foundation of the whole theory — the definition and construction of frames. Not just because of the existing linguistics research framework gives a rigorous conclusion, blog the sample single scenario is a simple sentence, frame semantics in complex sentence, such as metaphor, metaphors appear also have proved the excellent application, its limitation is that the logical structure and scenario analysis of clause.

However, as Google points out in the article, SLING will do well in corpus studies because frame semantics is based on corpora, so those who are new to NLP and those who like to make their own Chatbot should try it out and have a different experience later.

SLING making address: https://github.com/google/sling

The original address: https://research.googleblog.com/2017/11/sling-natural-language-frame-semantic.html

This article is compiled on wisdom, reprint please contact this public number for authorization.